Next, map the start and end positions of the answer to the original context by setting
   return_offset_mapping=True.